Docker Swarm is Docker's native cluster management solution. Swarm provides a native and deeper integration with Docker and exposes APIs that are compatible with Docker's remote APIs. Docker Swarm logically groups a pool of Docker hosts and manages them as a single large Docker virtual host. Instead of application administrators and developers deciding on which host the container is to be deployed in, this decision making will be delegated to Docker Swarm. Docker Swarm will decide which host to be used based on the bin packing and spread algorithms.

As Docker Swarm is based on Docker's remote APIs, its learning curve for those already using Docker is narrower compared to any other container orchestration tools. However, Docker Swarm is a relatively new product on the market, and it only supports Docker containers.

Docker Swarm works with the concepts of manager and nodes. A manager is the single point for administrations to interact and schedule the Docker containers for execution. Nodes are where Docker containers are deployed and run.

Kubernetes (k8s) comes from Google's engineering, is written in the Go language, and is battle-tested for large-scale deployments at Google. Similar to Swarm, Kubernetes helps manage containerized applications across a cluster of nodes. Kubernetes helps automate container deployments, scheduling, and the scalability of containers. Kubernetes supports a number of useful features out of the box, such as automatic progressive rollouts, versioned deployments, and container resiliency if containers fail due to some reason.

The Kubernetes architecture has the concepts of master, nodes, and pods. The master and nodes together form a Kubernetes cluster. The master node is responsible for allocating and managing workload across a number of nodes. Nodes are nothing but a VM or a physical machine. Nodes are further subsegmented as pods. A node can host multiple pods. One or more containers are grouped and executed inside a pod. Pods are also helpful in managing and deploying co-located services for efficiency. Kubernetes also supports the concept of labels as key-value pairs to query and find containers. Labels are user-defined parameters to tag certain types of nodes that execute a common type of workloads, such as frontend web servers. The services deployed on a cluster get a single IP/DNS to access the service.

Kubernetes has out-of-the-box support for Docker; however, the Kubernetes learning curve is steeper compared to Docker Swarm. RedHat offers commercial support for Kubernetes as part of its OpenShift platform.

Mesos is an open source framework originally developed by the University of California at Berkeley and is used by Twitter at scale. Twitter uses Mesos primarily to manage the large Hadoop ecosystem.

Mesos is slightly different from the previous solutions. Mesos is more of a resource manager that relays on other frameworks to manage workload execution. Mesos sits between the operating system and the application, providing a logical cluster of machines.

Mesos is a distributed system kernel that logically groups and virtualizes many computers to a single large machine. Mesos is capable of grouping a number of heterogeneous resources to a uniform resource cluster on which applications can be deployed. For these reasons, Mesos is also known as a tool to build a private cloud in a data center.

Mesos has the concepts of the master and slave nodes. Similar to the earlier solutions, master nodes are responsible for managing the cluster, whereas slaves run the workload. Mesos internally uses ZooKeeper for cluster coordination and storage. Mesos supports the concept of frameworks. These frameworks are responsible for scheduling and running noncontainerized applications and containers. Marathon, Chronos, and Aurora are popular frameworks for the scheduling and execution of applications. Netflix Fenzo is another open source Mesos framework. Interestingly, Kubernetes also can be used as a Mesos framework.

Marathon supports the Docker container as well as noncontainerized applications. Spring Boot can be directly configured in Marathon. Marathon provides a number of capabilities out of the box, such as supporting application dependencies, grouping applications to scale and upgrade services, starting and shutting down healthy and unhealthy instances, rolling out promotes, rolling back failed promotes, and so on.

Mesosphere offers commercial support for Mesos and Marathon as part of its DCOS platform.

Nomad from HashiCorp is another cluster management software. Nomad is a cluster management system that abstracts lower-level machine details and their locations. Nomad has a simpler architecture compared to the other solutions explored earlier. Nomad is also lightweight. Similar to other cluster management solutions, Nomad takes care of resource allocation and the execution of applications. Nomad also accepts user-specific constraints and allocates resources based on this.

Nomad has the concept of servers, in which all jobs are managed. One server acts as the leader, and others act as followers. Nomad has the concept of tasks, which is the smallest unit of work. Tasks are grouped into task groups. A task group has tasks that are to be executed in the same location. One or more task groups or tasks are managed as jobs.

Nomad supports many workloads, including Docker, out of the box. Nomad also supports deployments across data centers and is region and data center aware.

Fleet is a cluster management system from CoreOS. It runs on a lower level and works on top of systemd. Fleet can manage application dependencies and make sure that all the required services are running somewhere in the cluster. If a service fails, it restarts the service on another host. Affinity and constraint rules are possible to supply when allocating resources.

Fleet has the concepts of engine and agents. There is only one engine at any point in the cluster with multiple agents. Tasks are submitted to the engine and agent run these tasks on a cluster machine.

Fleet also supports Docker out of the box.